Automatic Image Annotation And Fast Similarity Search In Image Retrieval

With the rapid development of imaging technology, digital cameras and other imaging devices are becoming more and more popular. So the number of available images increases at an explosive speed. Further, the Internet greatly facilitates the communication between people. The exchange and deliverer of digital images are very cheap and convenient. Meanwhile, the ever increasing number of images brings problems to end users: they cannot find what they really need from huge amount of available data. Therefore, a lot of image retrieval and search technologies are developed.Present image retrieval usually depends on the annotation information, which is the textual description of an image. While the number of images is fast increasing, manually labeling all images becomes infesible. Therefore, automatic image annotation receives great attention and research effort in recent years. The most difficult problems are "semantic gap" and efficiency problems due to the huge number of images.Besides, content-based image retrieval (CBIR) is necessary in many application areas, such as medicial image retrieval. Automatic image annotation also needs to perform CBIR in many cases. The key problem in CBIR is to quickly and precisely find images similar to the query one. Because images are often represented as high-dimensional vectors and their huge amount, both the index and search are very difficult. When the number of images increases to millions or billions, such fast similar image search will be a very challenging research problem.This dissertation focuses on the automatic image annotation and fast similar image search. To make it clear, the main content and contribution are listed below:1. Introdutions to the automatic image annotation algorithms. The emphasis is put on the relevance-based models, generative models, and label propagation methods. Some recent research utilize the correlation between words, either statistical or semantic, to refine the image annotation. Some of this type of work is also discussed.2. This dissertation makes an analysis to the goal and available information in the automatic image annotation. Then, a unified annotation framework is proposed. The traditional annotation is extended to include two sub-problems: basic image annotation and annotation refinement. With the proposed framework, many previous annotation methods can be clearly undetstood.3. Based on the proposed framework, this dissertation presents several effective improved image annotation methods. These methods improve the image relation, word relation and lerning process, respectively. The experiments show the improvements are effective. It also helps validate the proposed annotation framework.4. This dissertation also discusses the similar image search. First, we restrict our focus on the detection of duplicate images within an image set. We propose an efficient and concise representation of an image. The proposed method has low computational complexity, needs little storage cost and can achieve high detection performance.5. The method in duplication image detection is further generalized to conduct similar image search. We propose to use multiple kinds of image features and exploit AdaBoost method to combine the concise representations of these features. So the similar image search in large image database can be quickly performed with good performance.